Speaker verification—an operational example

Abstract

An entry control system using speaker verification is now operational at Texas Instruments. This system has performed over 200 000 verifications during the past year and a half. Verification is based upon a 4-word utterance randomly constructed from a set of 16 reference words, each represented by filter-bank spectra. Voice prompting helps to stabilize the user's pronunciation. A sequential decision strategy provides good verification performance with an average of 1.3 utterances. Average verification time is less than 6 sec. Impostor acceptance is established at 1%. User rejection (0.2%) usually indicates anomalous user data. Speech reference data is adaptively updated at each successful entry attempt. Ten thousand bits of reference data storage are required for each user. Thorough user orientation is required to ensure successful enrollment. Enrollment lasts less than 2 min. [Development sponsored in part by the Air Force Systems Command, Rome Air Development Center and Electronics Systems Division.]